Computational Photography 15-862
Yajuan Wang

1. This is the result of a single-scale alignment result of lower resolution images. Sum of squared Differences (SSD) is used to find the vector that R and G channels should be moved to align with the B channel. The first column is the images after alignment and second column shows the result after cropping procedure. In all the result, the last two images seem that they can not align very well just based on the translation. If we want to get better result, further improvement should be input.


G: x=7, y=-1; R: x=10, y=-2

G: x=4, y=-1; R: x=9, y=1

G: x=1, y=-1; R: x=9, y=-2

G: x=7, y=-1; R: x=8, y=-2

G: x=5, y=0; R: x=11, y=-1

G: x=2, y=-1; R: x=10, y=-3

G: x=9, y=-1; R: x=10, y=-1

G: x=3, y=0; R: x=11, y=-1

G: x=15, y=2; R: x=11, y=3

G: x=-2, y=0; R: x=3, y=-2

2. This is the result of a multiscale pyramid result of high resolution images. Sum of squared Differences (SSD) is used to find the vector that R and G channels should be moved to align with the B channel. Here show the result after cropping.But it seems that result in the second line are not very well(I think they are not caused by translation among R G B images). The logic is that first I resize the image to 1/16 of the original one, and based on them find the vectors that they should move to align with each other. Then using the product of 16 and this vector to move the original images; then resize the moved original images to 1/8 size, find a new vector to align 1/8 images and so on.


G: x=71, y=24; R: x=147, y=47	G: x=41, y=-2; R: x=107, y=1


G: x=71, y=24; R: x=147, y=47	G: x=26, y=2; R: x=50, y=7	G: x=-18, y=3; R: x=48, y=-11

3. Here is the alignment results of other images I chose from previous work and all of them has been processed by cropping


G: x=6, y=0; R: x=10, y=-2

G: x=6, y=2; R: x=5, y=4